texar
texar copied to clipboard
Gumbel softmax trick with beam search
Is it possible to use Gumbel softmax trick with a beam search decoder in the existing implementation of Texar. I want to differentiate through the generated beam search samples into the generator network.
-
The current beam-search implementation does not support customization with decoder Helpers, though I think it's possible to refactor beam-search to enable the nice feature.
-
Here is a nice paper on direct optimization of beam search with continuous approximation: https://arxiv.org/pdf/1708.00111.pdf
Any contributions to 1) or 2) are welcome!